Runtime budgets: reserve bounded work before dispatching it

Track separate dimensions such as tokens, tool attempts, billed units, compute seconds, and wall-clock deadlines. They are not interchangeable without a declared conversion and price model. Count input tokens with the applicable tokenizer and include the prompt material actually sent. A generation output cap can bound one component only if the provider enforces it; hidden or separately billed work may require additional reserves or a softer policy. Label uncertain forecasts as estimates rather than pretending every vendor exposes a strict total-cost ceiling.

Tool retries, rejected attempts, deduplicated effects, cancellation, and cleanup can still consume resources. Decide whether a tool budget counts transport attempts, accepted executions, or logical effects, and record each where useful. A child run should receive a bounded allocation from its parent rather than independently treating the parent's whole budget as available. Lease expiry does not prove a remote request stopped, so reclaiming its reservation purely because a local timer fired can allow oversubscription.

Before dispatch, atomically reserve a request's declared resource vector if every dimension fits. Bind that reservation to logical run, child, step, attempt, and provider identity as needed by the accounting policy. Competing workers must not both approve against the same stale allowance. The budget transition and dispatch intent should share a supported durable protocol, with recovery deciding whether work was sent. A process-local counter or an unprotected read-then-subtract is not sufficient for a shared hard limit.

After confirmed completion, move actual usage into committed consumption and release the unused part of its reservation. Preserve the original reservation and settlement event so duplicate receipts do not charge or release twice. If actual cost exceeds the reservation, record the overrun and follow a defined containment policy; silently clipping the counter hides a broken bound. If the outcome is unknown, retain or transfer the obligation until authoritative usage or a safe upper-bound settlement is available. Cancellation is not a zero-cost receipt.

Run competing reservations that individually fit but jointly exceed a dimension, then verify only the allowed subset commits. Inject crashes after reservation, after dispatch, and before settlement. Test retries, duplicate usage receipts, child allocation returns, and cleanup spending. Report both estimates and measured use so systematic under-reservation becomes visible. Enforcement metrics should identify the exhausted dimension and remaining obligations rather than reduce everything to a generic too-many-steps error.

The function lab shows a declared synthetic cost family, not measured runtime usage. Reset, set a=1 and x=1, predict y=1, then Play or Step. Separately Reset, set a=1 and x=2, predict y=4, then Play or Step. In a hypothetical workload whose cost scales quadratically with input size, doubling size needs four cost units rather than two. The graph has no tokenizer, provider price, budget ledger, or upper-bound enforcement. Use it to check an explicit estimate's arithmetic, then validate the real bound and reservation protocol.